[Eeglablist] Choosing block size for Infomax ICA algorithm

[Makoto Miyakoshi]

Dear Yunhui and Jason,

> it seems that the two blocks produce quite different results. The
topography of top 10 major ICs changed, not just different order, or sign
flip.

That sounds interesting. I wonder if you can share your finding by putting
them online server accessible.

> Given that rejecting ICs can grossly changed the raw data, I wonder are
there any *scientific guide* to choose the block size?

A *scientific guide* is to change the parameter systematically as
independent variable to see the corresponding changes of the result as
dependent value. Every machine learning researcher does this to determine
parameters. If you want, you can do it. In doing so, you may want to use
getMIR() function, included in AMICA patckage, to evaluate how much mutual
information reduction is achieved by EEG.icaweight. How to use it was
demonstrated in Delorme et al. (2012) PLoS One paper. Anyway, the bottom
line for me is that it should not change the result drastically. So I agree
with Andreas.

AMICA has an option 'do_opt_block' to optimize the block size (which often
causes an error so I turn it off).

> Or is it a "dark area" where nobody bothers to explore?

At least for me it has been a 'dark area' since I have never explored the
parameters there. The best person to ask this question should be Jason
Palmer! Jason, if you have time, would you please give a comment?

Makoto

On Thu, Sep 27, 2018 at 9:08 AM Andreas Widmann <widmann at uni-leipzig.de>
wrote:

> Hi,
>
> Not really an answer to your question but to my understanding there is at
> least one misconception:
>
> >  keep the random number generator to a fixed seed, and it seems that the
> two blocks produce quite different results
> If you change block size you do actually reintroduce the randomness you
> intended to avoid by using a fixed seed. Data are shuffled first. Depending
> on block length different data enter the block level operations (adjustment
> of weights, computation of kurtosis etc.) resulting in different final
> weights. I would, however, not expect any *systematic* effects of
> increasing or decreasing block size but rather similar effects as if
> changing the seed (which can indeed be considerable but shouldn’t be
> systematic).
>
> Best,
> Andreas
>
> > Am 20.09.2018 um 16:51 schrieb 周云晖 <yhzhou17 at fudan.edu.cn>:
> >
> > Hi,
> >
> > When I checked the code of runica.m, I found that there are two choices
> of block size in the algorithm:
> >
> > % heuristic default - may need adjustment
> > %   for large or tiny data sets!
> > DEFAULT_BLOCK        = floor(sqrt(frames/4));  % heuristic default
> > DEFAULT_BLOCK          = ceil(min(5*log(frames),0.3*frames)); % heuristic
> >
> > and by default the second one is used. I have tried using the first one
> to a the sample dataset inside EEGLAB, and keep the random number generator
> to a fixed seed, and it seems that the two blocks produce quite different
> results. The topography of top 10 major ICs changed, not just different
> order, or sign flip. Given that rejecting ICs can grossly changed the raw
> data, I wonder are there any *scientific guide* to choose the block size?
> Or is it a "dark area" where nobody bothers to explore?
> >
> > The comments says different block size may suit different data size, but
> it is not clear whether I should use larger or smaller blocks for a larger
> dataset, and what kind of dataset can be considered "large".
> >
> > Best,
> >
> > Yunhui Zhou
> >
> >
> >
> > _______________________________________________
> > Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> > To unsubscribe, send an empty email to
> eeglablist-unsubscribe at sccn.ucsd.edu
> > For digest mode, send an email with the subject "set digest mime" to
> eeglablist-request at sccn.ucsd.edu
>
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to
> eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to
> eeglablist-request at sccn.ucsd.edu



-- 
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20180927/80d83eff/attachment.html>